Reproduce “Image Watermarks are Removable Using Controllable Regeneration from Clean Noise” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yepeng Liu, Yiren Song, Hai Ci, Yu Zhang, Haofan Wang, Mike Zheng Shou, Yuheng Bu
Researchers at the University of Florida, National University of Singapore, Tongji University, and InstantX Team developed CtrlRegen, a method that removes image watermarks, including robust high-perturbation types, by regenerating images from clean noise using a guided diffusion model. The technique incorporates semantic and spatial controls to preserve visual quality and consistency during watermark removal, demonstrating superior performance over existing regeneration attacks against various watermarking schemes.
Key insights
- Beginning the image regeneration process from clean Gaussian noise, rather than a partially noised latent, more effectively disrupts embedded watermark information, even for robust watermarks.
- Integrating explicit semantic and spatial control mechanisms into diffusion models enables high-quality image reconstruction while thoroughly removing watermarks, overcoming the trade-off between removal efficacy and image degradation.
- The 'no-box' nature of the attack allows it to be broadly applicable against various watermarking schemes without requiring specific knowledge of the watermarking method or detector.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.